{
  "id": 17551231,
  "title": "How much effort should a sub-agent use when thinking? CLAUDE_CODE_SUBAGENT_EFFORT from Claude Code 2.1.316 discusses the effort hierarchy of multi-agent coding.",
  "url": "https://urgent.news/2026/11/01/agent-claude-code-2-1-316-claude-code-subagent-effort-agent-effort",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-01T07:04:47.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/twinklemodel/zi-agent-gai-yong-duo-da-li-qi-xiang-cong-claude-code-21316-de-claudecodesubagenteffortliao-duo-agent-bian-ma-de-effort-fen-ceng-1poh"
  },
  "original_language": "zh",
  "account": "Anthropic has updated Claude Code to version 2.1.316, introducing a new setting called CLAUDE_CODE_SUBAGENT_EFFORT. This setting allows sub-agents to run at a specified effort level, rather than inheriting the main session's effort level. The update also includes changes to how effort levels are determined and used, with the goal of providing more granular control over the cost and performance of Claude Code. The company is moving towards making it possible to configure \"how much reasoning a given call should spend\" in a hierarchical and team-managed way.",
  "summary": "The short version summary (for short post channels) of Claude Code 2.1.316 released on October 31st adds CLAUDE_CODE_SUBAGENT_EFFORT: when no other settings specify, sub-agents run with this gear, no longer defaulting to inherit the main session's effort. This seems like a small switch, but it has a significant impact on the cost of multi-agent coding: the main session opens xhigh for planning, and sub-agents for searching, reading code, and running tests often use medium or even low. This article sorts out the complete priority of effort (environment variables > in-session choices > startup parameters > setting files > model defaults), how to write sub-agent frontmatter, and a set of layered configurations: \"high for main sessions, low for sub-tasks, and high for review alone\". When running slightly larger tasks with Claude Code, you will find that the main session does not actually use the most tokens, but rather the numerous sub-agents it dispatches.",
  "key_points": [
    "CLAUDECODESUBAGENTEFFORT setting determines sub-agent computing power",
    "Default medium effort recommended for most tasks to reduce cost"
  ],
  "editors_take": "Granular control over sub-agent computing power allows developers to optimize token usage and manage costs by aligning model effort levels with the specific complexity of each coding task.",
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}